Research Of Aerodynamic Data Modeling Based On Deep Neural Networks

Currently,we mankind are not only exploring the land and the sea,but has shifted the direction of exploration to the more vast sky and universe,thus posing a higher challenge to the technical strength of each country in the field of aerospace,and national research and progress in the field of aerodynamics has become particularly important.Aerodynam-ics field still relies on more traditional methods in obtaining aerodynamic data,but such methods are somehow limited and usually have a higher time or economic cost.How-ever,with the prosperous development of artificial intelligence,new opportunities have emerged,with which can lead to great advances in the way aerodynamic data are acquired in the field of aerodynamics.Traditional data modeling methods in aerodynamics require the use of mathematical methods to describe the physical laws and obtain a series of high-dimensional partial differential equations with complex structure,which are then solved by high-performance computers to acquire the data.The drawback of the methods above lies in the expensive cost in time.On the contrary,the efficiency of the models in data predicting can be greatly improved if deep learning methods are adopted away from complex physical mechanisms and focus on the mapping relationship between aerodynamic input parameters and their corresponding output data.Therefore,this thesis will explore how to build an aerody-namic data predicting model using deep neural networks,and establish the dataset used for experiments based on several representative cases in the field of aerodynamics to ver-ify the effectiveness of our aerodynamic data prediction model.The work of this thesis can be summarized as follows:1.I determined the experimental cases to be used in this thesis,and solved the experimental cases by computer according to the common methods in the industry,in order to construct the relevant data sets for the training of deep neural network models required in this thesis.2.After analyzing the relationship between input parameters and output data in aerodynamic data,a classical fully connected neural network was adopted and hyperparam-eters such as activation function,loss function,number of hidden layers and number of hidden layer nodes are determined to establish a prediction model for aerodynamic data,and the law of aerodynamic distribution in the aerodynamic data set was learned by the model.3.The model,named generative adversarial network(GAN),which have gained widespread attention in recent years,was investigated for exploring whether there are more novel and advanced deep neural network models that can be used for aerodynamic data modeling.Based on GAN,combined with the actual situation of the aerodynamic data modeling,this thesis adopts multilayer perceptrons instead of convolutional neural networks in the original GAN model,to construct the generative and discriminative net-works,and then,an aero-data predicting model was built.4.After completing the experiments related to the above two types of aero-data prediction models,I analyzed and summarized their advantages and shortcomings.Then,I adopted a method called multi-task learning and a neural network model called Cluster Network as improvements.The experimental results showed that the improved method was effective and can achieve more accurate prediction of aerodynamic data.